Neuroscience

Leveraging unlabelled data for generalizable neural population decoding

Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, Guillaume Lajoie

Featured July 20, 2026

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

By letting brain-reading models learn from lots of unlabeled brain signals, like filling in missing puzzle pieces, MOJO helps them understand brain activity better and predict actions more accurately, even with little labeled data.

In depth

The paper introduces MOJO, a novel training framework that combines self-supervised learning (SSL) with traditional supervised learning (SL) for spike-tokenizing neural decoders. By employing a masked autoencoder on latent neural representations, MOJO effectively leverages vast amounts of unlabelled neural data, significantly improving decoding performance, especially in scenarios with limited labelled data. This approach also leads to more interpretable neuronal representations that encode fine-grained unit properties and brain region information.

Key Takeaways

1
MOJO integrates self-supervised learning (SSL) via masked autoencoding with supervised learning (SL) to leverage unlabelled neural data, a critical advancement for label-impoverished settings.
2
The framework consistently achieves superior decoding performance across diverse neural datasets (spiking, ECoG) and species, outperforming purely SL-trained models, particularly in few-shot finetuning.
3
MOJO learns interpretable unit embeddings that accurately predict meta-features like brain region and spike statistics, demonstrating a deeper understanding of neural population dynamics without explicit optimization for these tasks.

Conceptual Flow

HIGH LEVEL

Methodology: Learning from All Brain Data

The new method teaches brain-reading models by letting them learn from both labeled examples (like 'this brain signal means moving an arm') and lots of unlabeled brain signals (like 'just listen to the brain activity').

Brain Signals

Movement Labels

Learn Patterns

Better Brain Model

Results: Smarter Predictions, Less Data

This new way makes the brain models much better at predicting actions, especially when they don't have many labeled examples, and helps them understand how different brain parts work together.

Old Brain Model

New Brain Model

Compare Performance

More Accurate

Needs Less Data

Understands Brain